deep-listeig
Deeplearigexperimetsforaudioclassificatio
Afullwrite-up,icludigtechicalexplaatiosaddesigdecisios,aswellasasummaryofresultsachievedcabefoudwithitheassociatedProjectReport.
ThisprojectcosistsofseveralJupyterotebooksthatimplemetdeeplearigaudioclassifiers.
1-us8k-ff-extract-explore.ipybthisotebookcotaiscodetoextractadvisualiseaudiofilesfromtheUrbaSoud8Kdatasetthefeatureextractioprocessusesaudioprocessigmetricsfromthelibrosalibrary,whichreduceseachrecordigto193datapoitsastheaudioiformatioishighlyabstracted,(wecaotprocesssuccessiveframesusigareceptivefield),thesefeaturesareitededtobefeditoafeed-forwardeuraletwork(FFN)2-us8k-ff-trai-predict.ipybthisotebookcotaisthecodetoloadpreviouslyextractedfeaturesadfeedthemitoa3-layerFFN,implemetedusigTesorflowadKerasalsoicludedissomecodetoevaluatemodelperformace,adtogeeratepredictiosfromidividualsamples,demostratighowatraiedmodelwouldbeusedtoidetifytheatureofliverecordigs3-us8k-c-extract-trai.ipybthisotebookextractsaudiofeaturessuitableforiputitoaclassic2-layerCovolutioalNeuralNetwork(CNN)muchmoreoftheaudiodataispreservedithisapproach,asthesavedumpyfeaturedataisover2GBIhave'ticludeditwiththisrepository,butyoucausethecodeithisotebooktoextractitfromtheorigialUrbaSoud8Kdataset4-us8k-c-salamo.ipybthisotebookimplemetsaalterativeCNN,similartooedescribedbySalamoadBello5-ffbird-c.ipybthisotebookusestheSalamoadBelloCNNtoprocesstheFreeField1010datasetoffieldrecordigs,withthegoalofrecogisigthepreseceofbirdsog.thedatasetisotpartofthisrepository,soifyouwattoruthiscodeyou'lleedtodowloadthedatayourself(seeistructiositheotebook)7-us8k-r-extract-trai.ipybthisusesaRecurretNeuralNetworktoclassifyMel-frequecycepstralcoefficiets(MFCC)features.Dogetitouchifyou'veayquestios,(me@jarocollis.com)
评论